Structural Operational Semantics for Stochastic Process Calculi Bartek Klin1 and Vladimiro Sassone2 1 Warsaw University, University of Edinburgh 2 ECS, University of Southampton Abstract. A syntactic framework called SGSOS, for defining well-behaved Mar- kovian stochastic transition systems, is introduced by analogy to the GSOS con- gruence format for nondeterministic processes. Stochastic bisimilarity is guaran- teed a congruence for systems defined by SGSOS rules. Associativity of parallel composition in stochastic process algebras is also studied within the framework. 1 Introduction Process algebras such as CCS [18] or CSP [5] are widely accepted as useful tools for compositional modeling of nondeterministic, communicating processes. Their se- mantics is usually described within the framework of Structural Operational Semantics (SOS) [19], where labelled nondeterministic transition systems (LTSs) are defined by induction on the syntactic structure of processes. Formalisms for SOS decriptions of nondeterministic systems have been widely studied and precisely defined (see [1] for a survey). In particular, several syntactic formats have been developed that guarantee certain desirable properties of the induced systems, most importantly that bisimulation is a congruence on them. Stochastic process algebras have been deployed for applications in performance evaluation, and more recently in systems biology, where the underpinning of labelled continuous time Markov chains (CTMCs), and more generally stochastic processes, is required rather than simple LTSs. Examples of such algebras include TIPP [11], PEPA [15], EMPA [3], and stochastic π-calculus [20]. Semantics of these calculi have been given by variants of the SOS approach. However, in contrast with the case of non- deterministic processes, SOS formalisms used here are not based on any general frame- work for operational descriptions of stochastic processes, and indeed differ substantially from one another. This is unfortunate, as such a framework would make languages eas- ier to understand, compare, and extend. Specifically, a format for SOS descriptions which guarantees the compositionality of stochastic bisimilarity, would make extend- ing process algebras with new operators a much simpler task, liberating the designer from the challenging and time-consuming task of proving congruence results. In this paper we define such a congruence format, which we call SGSOS. First we review existing approaches to the operational semantics of process algebras, concentrat- ing on the examples of PEPA [15] and the stochastic π-calculus [20]. As the operational techniques used there seem hard to extend to a general format for well-behaved stochas- tic specifications, we resolve to adapt a general theory of well-behaved SOS, based on category theory and developed by Turi and Plotkin [24]. The inspiration for our ap- proach comes directly from the work of F. Bartels [2], who used Turi and Plotkin’s results to design a congruence format for probabilistic transition systems. Standard operations of stochastic process algebras, as well as plenty of non-standard but potentially useful ones, fall within our format. Exceptions are recursive definitions and name-passing features of stochastic π-calculus, which we leave for future work. Within the SGSOS framework, we also investigate the issue of associativity of par- allel composition in stochastic process algebras, a design issue that, to our knowledge, has been overlooked in the literature. We notice in fact that in the original definition of stochastic π-calculus, parallel composition fails to be associative up to stochastic bisim- ilarity, and study conditions under which two forms of parallel composition, CSP-style synchronization and CCS-style communication, are associative. The structure of the paper is as follows. In §2 we recall previously studied ap- proaches to operational semantics of nondeterministic and stochastic systems. In §3 the bialgebraic theory of well-behaved SOS is recalled. In §4 we adapt the theory to obtain the SGSOS congruence format, with simple examples of GSOS specifications following in §5. The associativity of parallel composition is studied in §6, and in §7 we mention some directions of future work. Due to lack of space, all proofs are omitted in this extended abstract. 2 Transition systems and process calculi We begin our development by comparing two previously studied approaches to defining SOS for Markovian process algebras with the well-known world of SOS for nondeter- ministic systems such as CCS. 2.1 Nondeterministic systems and GSOS A labelled transition system (LTS) is a triple X, A, −→, with X a set of states, A a set a of labels and −→ ⊆ X × A × X a labelled transition relation, typically written x −→ y for (x, a, y) ∈ −→. An LTS is image-finite if for every x ∈ X and a ∈ A there are a only finitely many y ∈ X such that x −→ y. In the context of Structural Operational Semantics (SOS), LTS states are terms, and transition relations are defined inductively, by means of inference rules. For example, in a fragment of CCS [18], processes are terms over the grammar P ::= nil | a.P | P + P | P k P, and the LTS is induced from the following rules: a a x1 . y x2 . y a a a a.x . x x1+x2 . y x1+x2 . y (1) a a a a¯ x1 . y x2 . y x1 . y1 x2 . y2 a a τ x1kx2 . ykx2 x1kx2 . x1ky x1kx2 . y1ky2 Plenty of operators can be defined formally by rules like these. Indeed, the above speci- fication is an instance of a general framework for SOS definitions of LTSs (see e.g., [1]), called GSOS and defined formally as follows. An algebraic signature is a set Σ 3 f, g,... of operation symbols with an arity function ar : Σ → N, usually left implicit. The set of all terms over Σ with variables from set X is denoted TΣ X. In particular, TΣ 0 is the set of closed Σ-terms. 2 Fix a countably infinite set Ξ 3 x, y, z,... of variables. A GSOS inference rule [4] over a signature Σ and a set of labels A is an expression of the form n a j o n bl o xi . y j xi /. j 1≤ j≤k l 1≤i≤m c (2) f(x1,..., xn) . t where f ∈ Σ, n = ar(f), k, m ∈ N, i j, il ∈ {1,..., n}, a j, bl, c ∈ A, t ∈ TΣ Ξ, xi and y j ∈ Ξ are all distinct and no other variables occur in the term t. Expressions above the horizontal line in a GSOS rule are called its premises, and the expression below it is the conclusion.A GSOS specification is a set of GSOS rules; it is image-finite if it contains only finitely many rules for each f and c. Every GSOS specification Λ induces an LTS TΣ 0, A, −→ , with the transition rela- tion −→ defined by induction of the syntactic structure of the source states. For a term c s = f(s1,..., sn) ∈ TΣ 0, one adds a transition s −→ t for each substitution σ : Ξ → TΣ 0 such that for some rule r ∈ Λ as in (2), there is σxi = si, σt = t, and σ satisfies all a premises of r, meaning that for each premise x a . y there is σx −→ σy, and for each a a premise x /. there is no t ∈ TΣ 0 for which σx −→ t. An important property of the LTS induced by Λ is that bisimilarity on it is guaran- teed to be a congruence with respect to the syntactic structure of states. This means that GSOS is a congruence format for bisimilarity on LTSs. Moreover, it is easy to prove by induction that the LTS induced by an image-finite GSOS specification is image-finite. 2.2 Stochastic systems Just as nondeterministic process algebras are defined using labelled transition systems, the semantics of stochastic processes is often provided by labelled continuous time Markov chains (CTMCs). These are conveniently presented in terms of what we shall call rated transition systems (RTSs), i.e., triples (X, A, ρ), where X is a set of states, A + a set of labels and ρ : X × A × X → R0 is a rate function, equivalently presented as + an A-indexed family of R0 -valued matrices. The number ρ(x, a, y) is the parameter of an exponential probability distribution governing the duration of the transition of x to y with label a (for more information and intuition on CTMCs and their presentation by a transition rates see e.g. [12, 15, 20]). For the sake of readability we will write ρ(x −→ y) a,r a instead of ρ(x, a, y), and x −→ y will indicate that ρ(x −→ y) = r. The latter notation + suggests that RTSs can be seen as a special kind of A × R0 -labelled nondeterministic transition systems; more specifically, exactly those that are “rate-deterministic,” i.e., + a,r such that for each x, y ∈ X and a ∈ A there exists exactly one r ∈ R0 for which x −→ y. In the following we will consider image-finite processes, i.e. such that for each x ∈ X and a ∈ A there are only finitely many y ∈ X such that ρ(x, a, y) > 0. For such processes, the sum X a ρa(x) = ρ(x −→ y) (3) y∈X exists for each x ∈ X and a ∈ A; it will be called the apparent rate of label a in state x. a a Further, ρ(x −→ y)/ρa(x) is called the conditional probability of the transition x −→ y. It is the probability that x makes the transition provided that it makes some a-transition. 3 Various equivalence relations on states in RTSs have been considered. Of those, the most significant is stochastic bisimilarity (called strong equivalence in [14], and inspired by the notion of probabilistic bisimilarity from [17]), defined as follows.